International Journal of Computer Applications Technology and Research Volume 7–Issue 07, 281-285, 2018, ISSN:-2319–8656 Hangul Recognition Using Support Vector Machine Rahmatina Hidayati Moechammad Sarosa Panca Mudjirahardjo Department of Electrical Engineering Department of Electrical Engineering Department of Electrical Engineering University of Brawijaya State Polytechnic of Malang University of Brawijaya Malang, East Java, Indonesia Malang, East Java, Indonesia Malang, East Java, Indonesia Abstract: The recognition of Hangul Image is more difficult compared with that of Latin. It could be recognized from the structural arrangement. Hangul is arranged from two dimensions while Latin is only from the left to the right. The current research creates a system to convert Hangul image into Latin text in order to use it as a learning material on reading Hangul. In general, image recognition system is divided into three steps. The first step is preprocessing, which includes binarization, segmentation through connected component-labeling method, and thinning with Zhang Suen to decrease some pattern information. The second is receiving the feature from every single image, whose identification process is done through chain code method. The third is recognizing the process using Support Vector Machine (SVM) with some kernels. It works through letter image and Hangul word recognition. It consists of 34 letters, each of which has 15 different patterns. The whole patterns are 510, divided into 3 data scenarios. The highest result achieved is 94,7% using SVM kernel polynomial and radial basis function. The level of recognition result is influenced by many trained data. Whilst the recognition process of Hangul word applies to the type 2 Hangul word with 6 different patterns. The difference of these patterns appears from the change of the font type. The chosen fonts for data training are such as Batang, Dotum, Gaeul, Gulim, Malgun Gothic. Arial Unicode MS is used to test the data. The lowest accuracy is achieved through the use of SVM kernel radial basis function, which is 69%. The same result, 72 %, is given by the SVM kernel linear and polynomial. Keywords: Support Vector Machine; SVM; Kernel Polynomial; Kernel Linear; Kernel Radial Basis Function; Hangul 1. INTRODUCTION The next step, after pre-process, is feature extraction. It Optical Character Recognition (OCR) is a character has an important role on recognition process. It works through introduction system with images input. It contains texts that generating the basic component of the image called would be converted to the edited versions[1]. The work of features[9]. The feature extraction used in the current research OCR system depends on the kind of processed text. is chain code. The last process is recognition, using the SVM Generally, the text is divided into three categories. They are method with some kernels (linear, polynomial, and radial written, printed, and typed text[2]. basis function). Some researches on OCR System have been conducted. One of methods ever used is Support Vector Machine (SVM). 2. METHODOLOGY A kind of character which had ever been searched by using Generally, OCR system is divided into three main steps. SVM is Hindi number, which is known as Numeral Kanada, They are preprocessing (binarization, segmentation, and upper case and lower case alphabet A-Z [2,3,4]. The SVM thinning), feature extraction in which in this research uses method is used with different data, Korean characters known chain code, and recognition by applying Support Vector as Hangul. Machine (SVM) method. Figure 1 shows the general process Hangul recognition is more difficult compared with Latin of Hangul recognition. due to its complicated arrangement. Hangul is arranged from 2 dimensions (both left and right side), while Latin is arranged Hangul Image from left to the right [5]. A Research on Hangul recognition has ever been conducted, where the writer applies the Stochastic Relationship Modeling to the recognition process of Hangul Preprocessing syllable writing. The output of the research is Hangul syllable texts[6]. So far, the research on Hangul recognition is conducted Feature Extraction with Hangul texts output. The current research will improve the image conversion of Hangul with Latin text output. The image of Hangul converted into Latin text can be used as a learning material on how to read Hangul. Hangul Recognition OCR system, in general, is divided into three steps. They are preprocessing, feature extraction, and recognition. Pre- process includes three stages: binarization to change the Figure 1. Block diagram of Hangul recognition grayscale image into black white; segmentation, which is processed through connected component labeling, to separate the input into individual word; and thinning to decrease some The input image used in the current research is the letter information pattern (thin Line) in order to be easily analyzed and Hangul word. The letter consists of 21 vowels and 13 [7]. The research will employ algorithm Zhang Suen, which consonants, shown in Figure 2 with Latin letter. Each letter works faster than the other thinning algorithms[8]. has 15 different forms. The whole data are amounted to 510. www.ijcat.com 281 International Journal of Computer Applications Technology and Research Volume 7–Issue 07, 281-285, 2018, ISSN:-2319–8656 Figure 4. The example of Hangul word 2.1 Preprocessing The preprocessing includes 3 steps. First, the binarization or thresholding, is implemented to change the grayscale image become black white. The process of thresholding will produce the binary image, the image which has two steps of grayish (black and white). Generally, the process of floating the grayscale image to produce the biner image are as follows[11]: (1) With g (x,y) is the binary image from grayscale image f (x,y) and T assert the percentage of threshold. Second, thinning is used to decrease some information to a pattern becomes thin line in order to be easy to analyzed[7]. In this case, it will be applied the Zhang Suen algorithm which has faster performance compare with the other thinning algorithm[8]. To process thinning algorithm is shown in the Figure 2. The Letters of Hangul and Latin[10] Figure 6. This algorithm uses the pixel 3x3 and 8 degrees as in the Figure 5. P1 is a pixel that will be checked, if it fulfills Meanwhile for word, there are 6 ways on how to arrange the fixed condition, so the pixel will be deleted. The the letter of Hangul into word. The first type is shown in conditions are as follows[12]: Figure 3[10]. The discussion focuses on data type 2. (a) 2 ≤ B(P1) ≤ 6 (2) (b) A(P1) = 1 (3) (c) P2 x P4 x P6 = 0, and (4) ‘ (d) P4 x P6 x P8 = 0 (5) (e) P2 x P4 x P8 = 0, and (6) Figure 3. 6 the ways to arrange the letter of Hangul[10] (f) P2 x P6 x P8 = 0 (7) Meanwhile, the example of type 2 Hangul word is shown in Figure 4. Each word consists of 6 different forms. They are achieved by changing the word font. The used fonts are Arial, Batang, Dotum, Gaeul, Gulim, and Malgun Gothic. Figure 5. The pixel 3x3 with 8 degrees www.ijcat.com 282 International Journal of Computer Applications Technology and Research Volume 7–Issue 07, 281-285, 2018, ISSN:-2319–8656 Start Scan pixel 2 ≤ B(P ) ≤ 6 & Delete pixel Ya 1 A(P1) = 1 ? Tidak P2 x P4 x P8 = 0 & Delete pixel Ya P2 x P6 x P8 = 0 ? Tidak Figure 8. The direction of the sequence code in letter B P2 x P4 x P6 = 0 & Delete pixel Ya P4 x P6 x P8 = 0 ? In order to be processed into SVM, the feature extraction must have the same amount of features. Therefore, normalization needs to be done. It also aims to decrease the Tidak amount of feature which reoccurs. The normalization process for chain code can be done with the following pattern[12]: Matrix of thinning result (8) Note: N = the amount of feature wanted End = the amount of feature normalized = the amount of all letters normalized Figure 6. The Diagram algorithm of Zhang Suen 2.3 Support Vector Machine (SVM) The Support Vector Machine (SVM) is found in 1992 by 2.2 Feature Extraction Boser, Guyon, and Vapnik. The basic concept is the The next step after preprocessing is feature extraction, combination of computation theories introduced in previous which has an important role on recognition. This process will years, such as the margin hyperplane and kernel. The concept generate the necessary component of an image called of SVM can be seen as an effort to look for the best features[9]. The used feature extraction is chain code which hyperplane to separate two classes in the input grade. functions as the direction search. The direction usually uses Figure 9 shows some data of the two class member (+1 the following regulation Figure 7. and -1), where (a) shows some alternative separated line (Discrimination Boundaries), and (b) shows the best hyperplane. The best hyperplane can be found by measuring 3 2 1 the distance between the hyperplane and the nearer data 4 0 known as margin [13]. The data in the margin is called support vector[14]. 5 6 7 Figure 7. The direction of sequence code with 8 degrees[13] The sequence code is made through checking the direction of a pixel connected to the other pixels with 8 degrees. Each direction has different number. The pointed sign shows the first pixel. It is a place where the next steps will be fixed[13]. Figure 8 shows letter B with its sequence code. Figure 9. (a) The alternative separated field, (b) The best separated field with the biggest margin[14] www.ijcat.com 283 International Journal of Computer Applications Technology and Research Volume 7–Issue 07, 281-285, 2018, ISSN:-2319–8656 After achieving the slight feature, the system takes it in Practically, the real data does not only consist of two accordance with the direction search in Figure 7.